- Prompting techniques are methods of using pre-trained language models for natural language processing tasks without fine-tuning or adding new parameters
- They rely on designing natural language prompts and demonstrations that can elicit the desired output from the language model
Some of the prompting techniques are:
- Zero-Shot Prompting: no examples provided; leverage the model’s pre-training.
- Few-Shot Prompting: provide a few demos of input and output; show the desired reasoning format.
- Chain-of-Thought Prompting: prefix responses with intermediate reasoning steps; generate rich, concise summaries.
- Self-Consistency Prompting: pick the most frequent answer from multiple samples; increase redundancy and robustness.
- Tree-of-Thought Prompting: generate and evaluate multiple responses; allow backtracking through reasoning paths.
- Verifiers Prompting: train a separate model to evaluate responses; filter out incorrect responses.
- Fine-Tuning Prompting: fine-tune on an explanation dataset generated via prompting; improve the model’s reasoning abilities
- Prompting techniques can reduce the gap between pre-training and downstream tasks, and enable few-shot or zero-shot learning for new scenarios . Prompting techniques can also improve the performance, accuracy, and confidence of the language models on various tasks
- #NLP
- #Prompting
- #AI
- #LanguageModels